AI Research Roundup: High-Resolution Speech, Bioacoustics, and Model Interpretability

Breaking Down This Week's Most Impactful Papers in AI: From Earth Observation AI to Trustworthy Large Language Models

Contents

1. Wave-U-Mamba: An End-To-End Framework For High-Quality And Efficient Speech Super Resolution

2. BirdSet: A Large-Scale Dataset for Audio Classification in Avian Bioacoustics

3. Reinforcement Learning with Model Predictive Control for Highway Ramp Metering

4. Prithvi-EO-2.0: A Versatile Multi-Temporal Foundation Model for Earth Observation Applications

5. GIFT: A Framework for Global Interpretable Faithful Textual Explanations of Vision Classifiers

6. Understanding Model Calibration -- A gentle introduction and visual exploration of calibration and the expected calibration error (ECE)

7. E2Former: A Linear-time Efficient and Equivariant Transformer for Scalable Molecular Modeling

8. Tazza: Shuffling Neural Network Parameters for Secure and Private Federated Learning

9. DeTrigger: A Gradient-Centric Approach to Backdoor Attack Mitigation in Federated Learning

10. SELMA: A Speech-Enabled Language Model for Virtual Assistant Interactions

11. s1: Simple test-time scaling

12. Trust-Oriented Adaptive Guardrails for Large Language Models

13. Advances in Multimodal Adaptation and Generalization: From Traditional Approaches to Foundation Models

14. The "Huh?" Button: Improving Understanding in Educational Videos with Large Language Models

15. Emancipatory Information Retrieval

Wave-U-Mamba: An End-To-End Framework For High-Quality And Efficient Speech Super Resolution

Authors: Yongjoon Lee, Chanwoo Kim

Source and references: https://arxiv.org/abs/2409.09337v3

---

Introduction

This paper introduces Wave-U-Mamba, an end-to-end framework for high-quality and efficient speech super-resolution (SSR), which aims to enhance low-resolution speech signals by restoring missing high-frequency components.

Key Points

• Wave-U-Mamba directly performs SSR in the time domain, as opposed to conventional approaches that reconstruct log-mel features and then use a vocoder to generate high-resolution speech.